Anonymization of Gait Silhouette Video by Perturbing Its Phase and Shape Components

Abstract

Nowadays there are a lot of videos containing walking people on the web (e.g. YouTube). These videos can cause a privacy issue because the walking people can be identified by silhouette-based gait recognition systems which have been rapidly advanced in recent years. To solve the issue, in this paper, we propose a method for anonymizing human gait silhouettes. A gait silhouette consists of a static component including the body shape and a dynamic component including postures. We refer to the former and the latter as a shape component and a phase component, respectively. The proposed method anonymizes given gait silhouettes as follows: First, each of the given silhouettes is decomposed into its shape and phase components. Next, both components are separately perturbed. Finally, a new gait silhouette is generated from the perturbed components. Owing to the perturbation, the original silhouettes become less informative in the static aspect as well as the dynamic aspect, by which the gait recognition performance is seriously degraded. In our experimental results, the accuracy was actually degraded from 100% to 30% or less, without yielding any unnatural appearance in the output anonymized gait silhouettes.